Blog: Humans determine the success of machine learning
Algorithm and machine learning bias are hot topics causing optimism as well as concerns around fundamental shifts in technology. Tech is changing decision making across all aspects of life by finding patterns and learning from them. However it is humans who choose the data to generate these patterns from and interpret and use the results.
Alexandria Ocasio-Cortez made the news back in January talking about fixing the bias in algorithms causing racial profiling. There is the expectation of machine learning becoming the impartial decision maker unlike humans who have strong biases either based on their own beliefs or the institutions they represent. I agree with AOC’s position around fixing the bias however it feels like an unfair expectation from machine learning since decisions can only be as good as the data powering machine learning.
If you don’t have the rigor in managing training data forget about making unbiased decisions.
Amazon shelved their plans(for now) to run their recruiting on machine learning once they saw that their data was making the system biased against certain demographics. One can imagine that human recruiters had the same bias since they have been working on a similar dataset and making similar decisions.
Things get worse when incomplete datasets meet wrong incentives as we have seen in financial and government services. Navient agents have been accused of misguiding student loan borrowers to overpay for their repayments and not properly handling unique cases. In 2014, Governor of Maine introduced legislatures requiring deeper audits for TANF(Temporary Assistance for Needy Families) Families due to suspicions around misuse of the help, discouraging families from financial benefits that they’re entitled to. The decision was based on 0.3% of the 1.1million cash withdrawals(not transactions) made by the families.
Again the underlying problem was with the dataset and the people and their incentives who choose which datasets to use.
Luckily humans are still around to make sure that we have the right data set when making decisions.
Even if the algorithms and machine learning deliver results that seem biased against the consumer, humans with the right incentives can quality check the data and the decisions before they are executed. Especially those who are extremely close to the consumers like concierge services can collect new data on the spot and tweak the decision making if needed.
We should never underestimate the power of human interface. As long as we have the right incentives in place empathy is a very powerful tool to identify new data points needed for decision making. Conversation with a human is still the most familiar interface for consumers to share about their unique situation which would help complete the dataset.
Especially in industries like healthcare, government and financial services where the stakes are high, it is critical to invest in the human and leverage the power of conversation. And surely use machine learning to supercharge humans rather than replacing them.